Picture for Wojciech Samek

Wojciech Samek

Attribution-guided Pruning for Compression, Circuit Discovery, and Targeted Correction in LLMs

Add code
Jun 16, 2025
Viaarxiv icon

Deep Learning-based Multi Project InP Wafer Simulation for Unsupervised Surface Defect Detection

Add code
Jun 12, 2025
Viaarxiv icon

Relevance-driven Input Dropout: an Explanation-guided Regularization Technique

Add code
May 27, 2025
Viaarxiv icon

From What to How: Attributing CLIP's Latent Components Reveals Unexpected Semantic Reliance

Add code
May 26, 2025
Viaarxiv icon

The Atlas of In-Context Learning: How Attention Heads Shape In-Context Retrieval Augmentation

Add code
May 21, 2025
Viaarxiv icon

Steering CLIP's vision transformer with sparse autoencoders

Add code
Apr 11, 2025
Viaarxiv icon

Efficient Federated Learning Tiny Language Models for Mobile Network Feature Prediction

Add code
Apr 02, 2025
Viaarxiv icon

ASIDE: Architectural Separation of Instructions and Data in Language Models

Add code
Mar 13, 2025
Viaarxiv icon

Post-Hoc Concept Disentanglement: From Correlated to Isolated Concept Representations

Add code
Mar 07, 2025
Viaarxiv icon

FADE: Why Bad Descriptions Happen to Good Features

Add code
Feb 24, 2025
Viaarxiv icon